Ontological Smoothing for Relation Extraction

There is increasing interest in relation extraction, methods that convert natural language text into structured knowledge. The most successful techniques use supervised machine learning to generate extractors from sentences which have been labeled with the arguments of the relations of interest. Unfortunately, these methods require hundreds or thousands of training examples, which are expensive and time-consuming to produce. This paper presents ontological smoothing, a semi-supervised technique that exploits knowledge of a background ontology in order to learn extractors for a set of minimally-labeled relations. Ontological smoothing has three phases. The first step generates a mapping between the target relations and the background ontology. The second step uses knowledge-based weak supervision to heuristically generate new training examples for the target relations. The third step learns an extractor from a combination of the original and newly generated examples. Experiments on 61 relations across two target domains with Freebase as the background ontology show ontological smoothing can dramatically improve precision and recall.